Physical-AI teams source/capture thousands of hours of multimodal data and still don't know what it trains or if it's usable.
Euler answers that before the expensive run; raw recordings become evidence-backed, policy-ready datasets. Sharing our product glimpse ↓
Since moving to New York, my friend group has shifted from engineers to artists and creatives and never could I ever have imagined how much of a lifeline this would become for my soul amidst the madness that is tech
If you care about actually deploying robots, @RussTedrake's talk at CoRL was the most interesting one so far:
Three main things I took away:
1/ Stronger models need fewer interventions over the same deployment window, yet improve faster when retrained on that intervention data (see chart)
2/ They deploy as early as possible - even though that means that a remote teleoperator supervises the robot permanently.
3/ To reduce the robot-to-supervisor ratio, they train a separate head on the same policy that predicts when it needs help.
Today we’re open-sourcing Steel, our internal physics solver built with @NVIDIA.
It runs simulations up to 13× faster than PhysX while using ~3× less GPU memory.
Steel unlocks scaleable simulation for manipulation, for everyone.
Chennai could potentially be the next founder/startup hub, it’s ideal for any multi-disciplinary undertaking given it has all essential diverse industries at play with smooth public transit.
Lots of proactive work required to make it an attractive alternative though @CMOTamilnadu